Looksmaxxing for motion graphics. An agent skill that turns AI slop videos into films that look made by a motion designer.
The user's verdicts
What this user has actually said about films made for them (2026-09-28 → 2026-09-30), with context. Read it before you choose an idea. These are reactions to specific films, not universal rules; generalise them only as far as their context goes. When the user reacts to a new film, append an entry: date, film, quote, what they reacted to, and the narrowest lesson it supports.
Rejected
- Slide/PPT layouts are the strongest dislike. "This looks like a PPT, there is a text on left side, it's in structured." "Clear AI slop… the header, the footer, the text in left side, it is like presentation, it's like slides." Reacting to: header bars with product/version/year, left headline + grey subtitle, kicker above a headline, terracotta hairline, footer CTA with arrow circle.
- "Still slop, not perfect": a big claim on an empty flat ivory field, underline wipe, long holds, ivory/coral "editorial" palette.
- Copying the example instead of the brand. "It was just an example. Don't just take the reference from my site… its brand colour is orange, so it should put orange." The look comes from the subject's brand.
- Copying one reference. "The goal is not to copy/steal the recreation… take typography from this, gradient from another, effect from another… make something unique."
- Too sparse. "It should have more stuff than this, but it should use the animations, the fonts, the logos." Type alone on an empty ground was not enough; they wanted layered product, object and logo motion. (Richness ≠ clutter: see the calm-film pick below.)
- Vibe-coded objects: a "patch" drawn as a flat rounded yellow card with a dashed CSS-stitch corner was called slop; so was a dark rounded "Good morning / ● Rent in 3 days · $1,450" card. Objects and UI need real identity, not generic web-component styling.
Preferred
- Films where every joint had a physical cause (a click floods the frame, a caret becomes a number, the camera dives through a decimal) beat films of repeating cards.
- A film where one small motif carried through every beat of a specific, funny problem ("very good").
- Richer surfaces (photographic plates, a paper ground, a 3D object, real shutter blur) and designed sound (sectioned music, SFX, VO at −14 LUFS) beat flat-vector, silent versions: "Left is better in everyone."
- In one round the user picked a calm, legible film over a frantic, constantly moving one. Energy is not the goal; the right energy for the brand is.
- The user's standing goal: films that look "like an actual motion designer made them, like the references", killing AI slop without boxing the model's own creativity in.
Added 2026-10-05 (motionmaxxing v2)
Rejected
- The v1 Wispr Flow film (motionmaxxing v1 test). User-reported frame: tracked caps
WISPR FLOWtop-left, a chapter counter02 / SPEAKtop-right, a left-aligned headline on a flat dark ground. It is a slide/HUD template, the same anatomy as the three slop films the user rejected earlier (SuperX reel, "Claude motion reel", "muda": corner name, running counter or timecode, BPM/FPS text, left headline, card on a swatch). Caveat: an autopsy could not find that exact frame infilm/final.mp4(19.5 s), only its near-twin in the rejected "muda" film; treat the frame as the user's report and the film's other faults as measured. What the same film did, and why each is a fail:- Page chrome of any kind (brand label, counter, timecode, metadata, progress hairline) is not rescued by a reason. v1's law text said nothing about it, so the agent shipped label chips ("Messages", "Mail", "Editor") on every shot.
- Five flat swatch grounds in 19 s (cream, ink, cream, lavender, green) is a deck. v1 had prescribed "ground = brand's darkest or lightest surface, flipped by act". 0 of 9 captured media/screens were used.
- A 7 s hold on one generic rounded composer card (invented sender names, invented subject lines) is vibe-coded UI, 36% of the film.
- Headline typed letter by letter with a caret on an empty field: the rule (typing is for the product's input only) was clear and was rationalised away.
- A stat slide (centred 440 px "4x" over a small label) and a web-footer end card (logo, button, cursor).
- A pill pinned at one coordinate through five shots reads as a footer dock, not a carrier.
- 0.9 s of unchanged picture while a card waited; no NOTE.md was written, so nothing was self-reported.
- Lesson for v2: chrome, deck layouts and swatch grounds are mechanical fails (G5, "Never ship"); the picture is built from the captured material; every delivery has NOTE.md with measured numbers.
Self-reported, not a user verdict
- Resend v1 test (a private test run). The agent's own NOTE.md named the weakest spots: the emptiest stretch is a single small "Email"; the delivered card was 0.33 W for a 0.8 s read; two panels at 0.43 W; the music loop restarted; the end card ran about 5 s. That is the "small card on a void" failure (cf. the dark card on near-black the user called "vibe-coded" in the Kiln round, and the "too sparse" verdict above) and the "6 s static end card read as 'the film gave up'" lesson from the Astra round (see
edits.md). No user reaction is stored. Lessons: a proof object is magnified by the camera or cropped by the frame, it sits in a lit, textured world, and the end card holds 0.6-1.4 s.
Picks stored in the evals (for calibration; single rater)
- 09-29 sonnet-55: killing page chrome beat an ivory serif editorial with chrome and bar charts ("L is better").
- 09-30 spotify: one carried motif and a real material beat a left headline + phone; of three builds the one with real cover art won ("C is very good").
- 09-30 five invented brands, two models: 9 of 10 picks were with-skill; the one exception was a calm, legible film over a frantic one.
- 09-30 v2 (rich surfaces, designed sound) over v1 on all three reels: "Left is better in everyone."
- Pending: the 10-02 motion-eye A/B results (Spotify, OpenAI Dots) have no stored user pick. Ask before citing a winner; append it here when the user states it.
Standing instruction for appending
Date, film, quote, what they reacted to, narrowest lesson. Do not generalise beyond the film's context.